A performance optimization and stability guarantee method for a multi-computing layer real-time control system

Abstract

The application discloses a performance optimization and stability guarantee method of a multi-computing layer real-time control system, and relates to the fields of edge computing and real-time control. A new utility function, i.e., a time / state / utility function, is proposed for quantifying the relationship between control performance and system state. On this basis, a cross-time-scale task scheduling mechanism is designed: the edge server coordinates the scheduling strategy at a large time scale, and the local controller schedules the control task in real time at a small time scale. The scheduling mechanism divides the task into stability tasks and performance optimization tasks. The time / state / utility function is used to predict the task value, and the priority of the performance optimization task is dynamically adjusted to optimize the control performance. By designing a hard real-time stability task model, the stability guarantee problem is converted into a schedulability problem of the stability task. The application realizes online control performance optimization and stability guarantee of the multi-computing layer edge control system from the aspects of end-edge cooperation and real-time task scheduling.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of edge computing and real-time control, and particularly relates to a control performance optimization and stability guarantee method of a multi-computing layer real-time control system. BACKGROUND

[0002] With the maturity of edge computing and wireless technology, industrial automation is undergoing a major transformation. Traditional industrial automation relies on local controllers running on microcontrollers or programmable logic controllers, which are usually embedded in devices and wired to sensors and actuors. Due to the limited computing power of embedded devices, relying solely on local controllers cannot achieve monitoring, coordination, and human-machine interaction functions for a large number of physical controlled objects (such as distillation columns, turbines, machine tools, industrial robots, temperature control devices, etc.). In order to overcome the above limitations, industrial automation systems are developing into a multi-computing layer control system composed of edge layer and terminal layer. The terminal layer is composed of a large number of sensing terminals, local controllers and execution terminals, which are used to control multiple subsystems, while the edge server located in the edge layer usually has higher computing power and can play a role in monitoring and human-machine interaction. With the edge control system integrating more and more control loops and data-intensive computing tasks, it faces the following key challenges.

[0003] When multiple control tasks share a computing platform with limited capacity, meeting the hard real-time requirements of all control tasks will greatly limit the platform to accommodate a large number of control tasks. The difference between real-time scheduling performance and control performance further exacerbates this limitation. This difference is due to the complex relationship between actual control performance, execution of computing tasks and physical device state. At the same time, the changes in the physical state of the controlled object (such as temperature, speed, position, current, etc.) and the information subsystem state (such as end-to-end delay, etc.) that reflect the information processing process have different effects on device control. However, it is a key challenge to model the control performance of the information and physical state of the above control system in a closed form, even for scalar systems, which requires strong assumptions. In addition, there is an inherent conflict between stability and runtime control performance. Stability is usually guaranteed by offline design and analysis. In contrast, dynamic control performance changes at runtime, reflecting the transient process of device control. Therefore, dynamic control performance should be improved through online scheduling under the guarantee of stability. However, how to coordinate the conflicting factors in scheduling and control on multiple time scales is still a key problem.

[0004] In recent years, the control and scheduling joint design of real-time control systems can be mainly divided into three categories. The first category is offline scheduling, which guarantees static control performance. The second category is runtime scheduling without considering the physical state. However, the execution value of the control task changes with the physical state. In addition, it cannot adapt well to the physical disturbance at runtime because the scheduling strategy does not take the current physical state as a decision variable. The third category is runtime scheduling to guarantee online control performance. However, all studies ignore the delay, the physical state, and the specific relationship between runtime and static control performance.

[0005] Therefore, the present application is devoted to providing a performance optimization and stability guarantee method for a multi-computing layer real-time control system. First, the present application introduces a new utility function of the control task, namely the time / state / utility function, which is used to reflect the relationship between the real-time control performance and the information physical state. Subsequently, the present application proposes a multi-computing layer cross-time scale control task scheduling mechanism based on the time / state / utility function. Among them, the edge server coordinates the scheduling strategy in the large time scale to improve the control performance, and the local controller performs online real-time scheduling of the control task in the small time scale. The scheduling mechanism divides the control task into stability tasks for guaranteeing stability and performance optimization tasks for optimizing control performance; the scheduling scheme predicts the control value based on the time / state / utility function, dynamically adjusts the priority of the performance optimization task, and realizes online optimization of the control performance; at the same time, the system stability guarantee problem is converted into the schedulability problem of the stability task, and a hard real-time stability task model is designed to obtain the guarantee of stability. The present application realizes online control performance optimization and stability guarantee of the multi-computing layer edge control system from the aspects of end-edge cooperation and real-time computing task scheduling. SUMMARY

[0006] In view of the above defects of the prior art, the technical problem to be solved by the present application is to realize control performance optimization and stability guarantee of a multi-computing layer edge real-time control system.

[0007] To achieve the above purpose, the present application provides a performance optimization and stability guarantee method for a multi-computing layer real-time control system, comprising the following steps:

[0008] Step 1, establishing a data-driven time / state / utility function;

[0009] Step 2, analyzing the stability condition of the control loop;

[0010] Step 3, the edge server performs priority prediction of the control task in the large time scale;

[0011] Step 4, the local controller performs real-time scheduling of the control task based on the priority in the small time scale.

[0012] Further, the time / state / utility function represents the control value, and builds a coupling model between the delay, the physical state and the control performance.

[0013] Further, the step 1 comprises the following steps:

[0014] Step 1.1, defining the input and output variables of the time / state / utility function;

[0015] Further, the input variables of the machine learning model include the information of the real-time control system and the physical state information, including the physical controlled object parameters, the physical controlled object state, the time delay, etc.; and the output variable is the performance of the control system.

[0016] Step 1.2, collecting the input and output data of the time / state / utility function and dividing the training set and the test set;

[0017] Step 1.3, selecting a machine learning model, training the model and testing the model accuracy.

[0018] Further, the machine learning model includes a polynomial regression model, a support vector machine, a least squares boosting method and a neural network, etc.

[0019] Further, the model accuracy testing method includes calculating the normalized error of the model and calculating the absolute error of the model, etc.

[0020] Further, the step 2, the stability of the control loop, is represented by the schedulability of the real-time control task triggered at a stable period.

[0021] Further, the step 2 comprises the following steps:

[0022] Step 2.1, finding the stable period, that is, the maximum control period that ensures the stability of the control loop;

[0023] Step 2.2, ensuring the schedulability of the real-time control task triggered at a stable period.

[0024] Further, the step 3 comprises the following steps:

[0025] Step 3.1, dividing the stability task and the performance optimization task;

[0026] Step 3.2, predicting the scheduling priority of each performance optimization task;

[0027] Step 3.3, explicitly assigning the priority of the stability task and the performance optimization task.

[0028] Further, the step 3.2, the edge server predicts the scheduling priority of each performance optimization task according to the time / state / utility function and the physical state.

[0029] Further, the priority of all stability tasks is higher than the priority of performance optimization tasks in the step 3.3. In this way, the stability tasks can be guaranteed to be schedulable, i.e. the stability of the control system.

[0030] Further, the edge server collects the state of the physical controlled object and the execution state of the real-time control task from the local device at a large time scale, predicts the value of the control task based on the real-time information physical state, and dynamically adjusts the priority of the control task according to the prediction result to coordinate the global resource allocation.

[0031] Further, the local controller completes the real-time scheduling of the control task online based on the priority provided by the edge server.

[0032] The application belongs to the field of edge computing and real-time control, and mainly relates to a control task real-time scheduling mechanism for a multi-computing layer real-time edge control system, in particular to a control task real-time scheduling mechanism based on control value aiming to realize control performance optimization and stability guarantee.

[0033] The existing scheduling scheme adopts offline static or online scheduling strategy without considering the physical state, which cannot adapt to various disturbances and other phenomena of the physical controlled object in runtime, thereby causing significant decline of the control performance of industrial equipment. The application designs an edge control architecture, and the edge server adjusts the priority of the local control task according to the state of the physical controlled object and the execution state of the real-time control task at each large time scale online, and then responds to the state change of the physical equipment in time. The multi-computing level edge real-time control system architecture combines the large time scale coordination ability of the edge server and the fine-grained real-time scheduling ability of the local controller, and dynamically adapts to the disturbance in the operation of the physical system. (1) The coordination effect of the edge server: the edge server collects the running state of the physical system from the local device in a larger time scale, and predicts the value of the control task based on the real-time information physical state. According to the prediction result, the priority of each control task is dynamically adjusted, and the global resource allocation of the task is coordinated. (2) Fine-grained scheduling of the local controller: the local controller completes the real-time scheduling of the control task online based on the priority provided by the edge server. The application (1) improves the real-time performance and dynamic adaptability of the system: by adjusting the task priority online, the scheduling scheme can dynamically adapt to the state change of the physical equipment, thereby significantly improving the response ability of the system to physical disturbance; (2) optimizes the resource utilization efficiency: resources are preferentially allocated at the critical physical state change (such as external disturbance), avoiding the blindness of resource allocation, reducing the performance decline caused by resource waste, and improving the overall system efficiency; (3) simplifies the scheduling design in complex scenarios: the dynamic scheduling of the edge server reduces the preset demand for various extreme situations in the offline design process, reduces the design complexity, and enhances the universality and expandability of the system; (4) improves the stability and reliability of the system: by linking the physical state and the scheduling decision, the risk of disconnection between the scheduling scheme and the device state is reduced, and the system can still run stably under complex interference conditions.

[0034] The existing online scheduling strategy ignores the complex coupling relationship between construction delay, physical state and runtime and static control performance. There are limitations in modeling the complex coupling relationship using analytical expressions, such as the current technology only supports specialized models such as vector systems or low-dimensional linear systems. The present application adopts a data-driven method to obtain time / state / efficiency functions to represent control value, which relates to calculation delay, physical state and runtime and static control performance. The present application accurately describes the coupling relationship between calculation delay, physical state and control performance by constructing a data-driven control value prediction model, and designs an online scheduling strategy based on this. The principles mainly include the following three points: (1) comprehensive representation of control value: introduce time / state / efficiency functions, and based on data-driven methods, associate the following key factors: ① calculation delay ② physical state ③ control performance. (2) Coupling relationship modeling and optimization: convert the above factors into quantifiable indicators through the control value function, and construct the coupling relationship. In the scheduling process, the coupling model is used to realize the dynamic allocation of task priority. (3) Data-driven prediction ability: use real-time information physical state to dynamically predict the control value, ensuring that the scheduling scheme can be adjusted in a timely manner in response to runtime changes. The present application (1) accurately represents the complex coupling relationship: using a data-driven method, a coupling model of calculation delay, physical state and control performance is constructed, which accurately describes the behavior of complex dynamic systems. A unified time / state / efficiency function framework is provided, which associates multiple performance indicators under the same optimization objective, significantly reducing the complexity of the system scheduling strategy; (2) improve the real-time performance and flexibility of the system: dynamically update the efficiency function to ensure that the scheduling strategy can quickly adapt to changes in physical state and task delay, significantly improving the response speed and flexibility of the system; (3) reduce design and scheduling costs: the data-driven method reduces the dependence on complex mathematical derivations in traditional model construction, reducing design costs and improving the practicality of scheduling algorithms in complex systems; (4) improve control performance: real-time calculation delay and control performance are directly introduced into the scheduling decision, avoiding the disconnection between the scheduling process and the system control target. Control performance optimization can be achieved in different operating scenarios, including the balance of transient and steady-state performance.

[0035] Most of the existing real-time scheduling schemes emphasize that all task instances must meet the schedulability requirement, resulting in extremely pessimistic resource utilization. The present application proposes a task scheduling method based on control value prediction, which distinguishes between stability tasks and performance optimization tasks, and different categories of tasks have different schedulability requirements and scheduling methods. The present application breaks through the limitations of traditional scheduling methods by introducing a classification mechanism for control tasks and a differentiated scheduling strategy based on control value prediction, effectively improving resource utilization.(1) Task classification and differentiated scheduling: ① Stability tasks: key tasks to ensure system stability, must meet strict hard real-time schedulability requirements; ② Performance optimization tasks: to improve the control performance of the system, use the remaining computing resources for scheduling, priority can be dynamically adjusted; (2) Dynamic priority adjustment: based on the control value prediction model, dynamically evaluate the scheduling value of performance optimization tasks, and adjust their priority according to the real-time situation of system resources. Ensure the high priority scheduling of stability tasks, while executing performance optimization tasks under the remaining resources; (3) Transformation of stability problem: transform the problem of guaranteeing system stability into the schedulability problem of stability tasks, and use the theoretical hard real-time task model to provide strict guarantee for stability. The present application(1) improves resource utilization: by dividing tasks into stability tasks and performance optimization tasks, the present application allows performance optimization tasks to moderately sacrifice their scheduling requirements when resources are scarce, thus avoiding excessive conservatism in resource allocation; (2) Enhance the scheduling flexibility of the system: based on the control value prediction method, the priority of different tasks is dynamically adjusted, making the scheduling strategy flexible to adapt to changes in the running environment and task requirements; (3) Guarantee system stability: strict schedulability of stability tasks ensures system stability; (4) Suitable for complex real-time system scenarios: the scheduling method of the present application is particularly suitable for complex real-time systems with a large number of tasks and large differences in task characteristics, and it performs superior scalability and applicability in a multi-task environment.

[0036] Compared with the prior art, the present application has the following obvious substantial features and significant advantages:

[0037] 1. Technical advantages of the present application

[0038] The control task real-time scheduling strategy based on control value prediction proposed by the present application provides an innovative solution for multi-computing level edge real-time control systems, overcoming many limitations of traditional industrial automation systems in dynamic control performance optimization and stability guarantee. The specific advantages include:

[0039] 1.1 Strong dynamic adaptability: through the coordination of control task priority by the edge server on a large time scale and the online scheduling of the local controller, it can respond to physical device state changes in real time, effectively dealing with runtime disturbances and dynamic demand changes.

[0040] 1.2 Resource utilization optimization: Fine-grained division of control tasks into stability tasks and performance optimization tasks, achieving hierarchical efficient utilization of computing resources, overcoming the problem of resource waste caused by hard real-time constraints in traditional scheduling schemes.

[0041] 1.3 Theoretical stability guarantee: By converting the stability guarantee problem into the schedulability problem of stability tasks, and designing a hard real-time stability task model, the system stability is theoretically guaranteed.

[0042] 1.4 Data-driven control performance optimization: Using data-driven methods to build control value prediction models, dynamically adjusting the priority of performance optimization tasks online, significantly improving real-time control performance.

[0043] 2. Performance indicators of the present application

[0044] According to the current technical evaluation and experimental results, the present application shows significant advantages in the following key performance indicators:

[0045] 2.1 Real-time guarantee of stability tasks: In complex industrial environments, since stability tasks need to meet hard real-time constraints, the completion rate of stability tasks always remains 100%, which successfully guarantees the stability of the system.

[0046] 2.2 Dynamic control performance: After optimization based on control value prediction, the performance indicators of the control system such as mean absolute error are improved by more than 90% compared with the scheduling results based on rate monotonic scheduling strategy.

[0047] 3. Production implementation of the present application

[0048] The technical solution of the present application has good industrial adaptability and is easy to combine with the hardware and software architecture of existing industrial automation systems.

[0049] 3.1 Hardware implementation:

[0050] 3.1.1 The hardware platform required by the edge server (such as industrial server, edge computing node) has been widely used in industrial scenarios, and no major modification is needed for existing equipment.

[0051] 3.1.2 The local controller can be compatible with multiple embedded computing platforms, and only needs to update the software algorithm to complete the upgrade and deployment.

[0052] 3.2 Software integration:

[0053] 3.2.1 The scheduling algorithm of the present application supports standard industrial wired communication protocols (such as EtherCAT, Profinet, etc.), which is convenient for rapid integration into existing control systems.

[0054] 3.2.2 The control value prediction model based on data-driven control can be trained and deployed using existing industrial big data platforms, significantly reducing model development costs.

[0055] 4. Industrial application prospects of the present application

[0056] 4.1 Industrial automation scenarios:

[0057] The present application has wide application prospects in intelligent manufacturing, industrial robots, process control and other scenarios, especially in complex industrial environments that require dynamic optimization of control performance, which can significantly improve the flexibility and intelligence level of the system.

[0058] 4.2 Smart grid and energy management:

[0059] The real-time scheduling characteristics of the multi-computing-level edge control system are suitable for application in distributed energy management in smart grids, optimizing resource scheduling and ensuring system stability.

[0060] 4.3 Transportation and autonomous driving:

[0061] In the transportation system with high demand for multi-task concurrency and dynamic scheduling, the technical solution of the present application can improve real-time performance and stability, providing technical support for intelligent transportation and the Internet of Vehicles.

[0062] The concept, specific structure and technical effects of the present application will be further described below in conjunction with the accompanying drawings, so as to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a double-computing-layer architecture diagram of an edge control system of a preferred embodiment of the present application;

[0064] Figure 2 is a machine learning model of a time / state / utility function of a preferred embodiment of the present application;

[0065] Figure 3 is a classification and release mode of stability and performance optimization tasks of a preferred embodiment of the present application;

[0066] Figure 4 is a control task scheduling mechanism based on control value prediction of a preferred embodiment of the present application. DETAILED DESCRIPTION

[0067] The following reference drawings introduce several preferred embodiments of the present application, making its technical content clearer and easier to understand. The present application can be embodied in many different forms of embodiments, and the scope of protection of the present application is not limited to the embodiments mentioned herein.

[0068] In the drawings, components having the same function are denoted with the same reference numerals, and components similar in structure or function are denoted with similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the size and thickness of each component are not limited in the present application. In order to make the drawings clearer, the thickness of some components is appropriately exaggerated in some places in the drawings.

[0069] The application provides a task real-time scheduling strategy based on control value prediction for a multi-computing layer real-time edge control system, which can realize online optimization of control performance and stability guarantee.

[0070] As Figure 1 shown, the technical scheme adopted by the application to solve the technical problem is: a control task real-time scheduling mechanism based on time / state / utility function, which aims to realize control performance optimization and stability guarantee, containing the following steps:

[0071] Step 1: Establish a data-driven time / state / utility function;

[0072] Considering the control system model shown in formula (1):

[0073]

[0074] Where i∈{1,2,...,M} is the control loop index, k is the index of control period T i , x i (k) is the physical state vector, u i (k) is the control instruction vector, θ i is the configurable model parameter vector, f i (·) and g i (·) are the device dynamics model and control law respectively. The physical state x i (k) of the device is collected by the sensor and sent to the local controller at a period T i . The local controller generates and schedules the execution command u i (k) of the control device for the actuator according to x i (k) and the scheduling priority decided by the edge server within each period T i . M control tasks share the local processor, and each control task τ i has an implicit deadline D i =T i .

[0075] The offline training of the machine learning model obtains the time / state / utility function , and further obtains the relationship between V i and the information physical state:

[0076]

[0077] 1.1) Define the input and output variables of the time / state / utility function;

[0078] The input variables of the time / state / utility function are s i , and x i , where the static properties s i remain constant throughout the operation of the real-time control system, including the control period T i , the model parameters θ i , and the control policy g i (·), etc.; the cyber-physical state includes the end-to-end delay and the physical state x i , etc. The output variable V i represents the prediction of the control performance (e.g., the mean absolute error, the maximum absolute error, and the settling time).

[0079] 1.2) Collect the input and output data of the time / state / utility function and divide the training set and the test set;

[0080] The simulation of the control system is performed by sampling the input values (the end-to-end delay the physical state x i ) within the feasible region in the cyber-physical state. The collected data are in the format where h i is the prediction horizon, which is set depending on the time constant of the control system. After the collection of the data is completed, the training data set and the test data set are divided in a certain proportion.

[0081] 1.3) Select a suitable machine learning model, perform model training, and test the model accuracy.

[0082] The machine learning model, such as the polynomial regression model, the support vector machine (SVM), the least squares boosting method (LSBoost), or the neural network (NN), is used to perform model training on the training data set, as shown in Figure 2 . The trained model is tested for accuracy on the test data set, such as calculating the normalized error or the absolute error of the model, etc.

[0083] Step 2: Analyze the stability conditions of all control loops;

[0084] 2.1) Find the stable period, i.e., the maximum control period that ensures the stability of the control loop;

[0085] The stability of the control loop can be ensured by the maximum control period, i.e., the stable period SPi to characterize. Stability period SP i is a multiple of control period T i , indicating that the system can still guarantee stability when facing up to n i maximum consecutive deadline miss behaviors.

[0086] Calculate stability period SP i : the stability condition of linear time-invariant system studied by weak hard task model, get the control task τ i The maximum number of consecutive deadline misses n i under the premise of guaranteeing stability, stability period SP i is calculated by formula (3):

[0087] SP i = (n i + 1)T i (3)

[0088] Where n i is the maximum value that satisfies the joint spectral radius ρ(∑ i )<1, is the closed-loop characteristic matrix of loop i when the control task τ i satisfies the deadline constraint, A o is the closed-loop characteristic matrix of loop i when the control task τ i misses the deadline.

[0089] 2.2) Ensure the schedulability of real-time control tasks triggered at a stable period.

[0090] The schedulability of stability tasks can be tested by the analysis algorithm corresponding to the scheduling strategy. For example, if the deadline monotonic algorithm is used, the worst-case response time can be verified by response time analysis to see if it exceeds the deadline; if the earliest deadline first algorithm is used, the total execution time demand of tasks in any time period can be verified by processor demand analysis to see if it exceeds the processor available time.

[0091] Step 3: The edge server performs priority prediction of control tasks at a large time scale;

[0092] 3.1) Divide stability tasks and performance optimization tasks;

[0093] The first task within each stability period SP i of the task τ i is designated as a stability task, and the jth stability task during the entire running period is denoted as The period of stability task is SP i , and the constraint deadline is T i . τi The remaining tasks other than stability tasks are defined as performance optimization tasks, denoted as The task release pattern of both is shown as Figure 3 .

[0094] 3.2) Predict the scheduling priority of each performance optimization task;

[0095] At each runtime coordination cycle, the edge server runs a time / state / utility function and runtime information-physical state to predict the scheduling priority (control value) of each performance optimization task.

[0096] 3.2.1) At the beginning of each coordination cycle T c , the edge server calculates the control value density δ c (k') for the period [k'T c , (k+1)T i ] according to the time / state / utility function.

[0097]

[0098] where is the time / state / utility function, C i is the worst-case execution time of control task τ i , and k' is the time index of large time scale priority prediction. Δ is the end-to-end delay interval for control value density prediction. Since δ i and Δ are coupled, in order to decouple scheduling and value density prediction to derive δ i , the same end-to-end delay interval Δ is applied to all control loops for fairness. In this way, δ i represents the execution of under the given runtime physical state x i (k' ) ) at the sensitivity of the calculation delay.

[0099] 3.2.2) Set the priority set P(k') = [p1(k'), p2(k'),..., p M (k')] of M control loops according to the ranking of value density δ i (k').

[0100] 3.3) Explicitly assign the priority of stability tasks as well as performance optimization tasks, as shown in Figure 4 .

[0101] 3.3.1) Assign the highest priority range [1, M] to the M stability tasks

[0102] Assign the M highest priority [1,M] values ​​to the M tasks τ. i Stability tasks This allocation is not limited to any priority allocation strategy, such as monotonic application deadline rate or earliest deadline priority.

[0103] 3.3.2) Assign the priority range [M+1, 2M] to the M performance optimization tasks.

[0104] For performance optimization tasks Map the priority set P(k') of the performance optimization tasks obtained in step 3.2.2) within the range [M+1, 2M].

[0105] 3.3.3) The priorities of stability tasks and performance optimization tasks are transmitted to the local controller via the edge wired network.

[0106] Step 4: The local controller performs priority-based real-time scheduling of control tasks on a small time scale.

[0107] The local controller executes a priority-based control task scheduling strategy, with priorities transmitted from the edge server to the local controller at the beginning of each large time scale.

[0108] The local controller uses the P provided by the edge server i (k') in the next coordination period T c Real-time scheduling based on this priority is performed within the system. Until an update is received, the priority remains the same as the previous scheduling priority P. i (k'-1) are the same.

[0109] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for performance optimization and stability assurance of a multi-computation-layer real-time control system, characterized in that, Includes the following steps: Step 1: Establish a data-driven time / state / utility function; Step 2: Analyze the stability conditions of the control loop; Step 3: Edge servers perform priority prediction of control tasks over a large time scale; 3.1) Divide the task into stability tasks and performance optimization tasks; Task Each stable cycle The first task within the entire runtime is designated as a stability task, and the [number]th task during the entire runtime... A stability task is represented as Stability task The period is The constraint deadline is ; In addition to stability tasks The remaining tasks are defined as performance optimization tasks, denoted as... ; 3.2) Predict the scheduling priority of each performance optimization task; 3.2.1) Edge servers in each coordination cycle At the start time, the period is calculated based on the time / state / utility function. Internal control value density : in For time / state / utility functions, It is a control task Worst-case execution time It is a time index for priority prediction on a large time scale. It is the end-to-end delay interval used to control value density prediction. Represents the physical state at a given runtime. Execute Sensitivity to computational latency; 3.2.2) Based on value density Sorting settings Priority set of each control loop ; 3.3) Priority allocation for stability tasks and performance optimization tasks; 3.3.1) Set the highest priority range Assigned to A stability task; 3.3.2) Priority range Assigned to One performance optimization task; 3.3.3) Transmit the priority of stability tasks and performance optimization tasks to the local controller; Step 4: The local controller performs priority-based real-time scheduling of control tasks on a small time scale; The local controller executes a priority-based control task scheduling strategy, with priorities transmitted from the edge server to the local controller at the beginning of each large time scale.

2. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, characterized in that, The time / state / utility function characterizes the control value and constructs a coupled model between computational delay, physical state, and control performance.

3. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Define the input and output variables for the time / state / utility function; Step 1.2: Collect the input and output data of the time / state / utility function and divide it into training and test sets; Step 1.3: Select a machine learning model, train the model, and test its accuracy.

4. The performance optimization and stability assurance method for a multi-computation layer real-time control system as described in claim 3, characterized in that, in step 1.1, the input variables of the machine learning model include the information and physical state of the real-time control system, wherein the physical state includes the parameters of the physical controlled object and the state of the physical controlled object, and the information state includes the time delay; the output variable is the performance of the control system.

5. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, characterized in that, In step 2, the stability of the control loop is characterized by the schedulability of real-time control tasks triggered by a stable cycle.

6. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Determine the stable period, which is the maximum control period that guarantees the stability of the control loop; Step 2.2: Analyze the schedulability of real-time control tasks triggered by a stable cycle.

7. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, wherein step 3.2 is characterized in that, The edge server predicts the scheduling priority of each performance optimization task based on time / state / utility functions and physical state.

8. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, wherein step 3.3 is characterized in that, All stability tasks have a higher priority than performance optimization tasks.

9. The method for performance optimization and stability assurance of a multi-computation layer real-time control system as described in claim 1, characterized in that, The edge server collects the state of the physically controlled object and the execution status of the real-time control task from the local device on a large time scale. Based on the real-time information of the physical state, it predicts the value of the control task according to the time / state / utility function. Based on the prediction results, it dynamically adjusts the priority of the control task and coordinates the real-time scheduling of the computing tasks on the local controller.

Images